<mastermart>
I think I was too paranoid, about scheduling the holes, the stuff can be authored in many ways, but scheduling backwards randomly would work too.
<mastermart>
cause it tries to schedule the holes only when registers are ready, and the feed is parallel there, no matter what stete it is , it will still try that.
<mastermart>
they name such method to do whatever needed as soft-clauses, that is a dependent load on vector elements, which replays one element more than once.
<mastermart>
since dispatcher sits in fsm, poking for writebacks to increment the fetch whid, it will runahead of simd engines arbiter that way.
<mastermart>
The FSM is idle execute resultsready kind of stuff.
<mastermart>
if any of the instructions are replayd more than once, it will happen
<mastermart>
but this can not happen with alus only LSU loads.
marvs has joined #lima
megi has joined #lima
<mastermart>
so the CUBLAS, LAPACK, FLAME,FLASH supermatrix slides seem to be correct, if i managed to read them myself right. I am not sure why the benchmark did not get the whole single precision max throughput though:(
<mastermart>
but actually this even makes sense, since the whole throughput FLOPS comes when you have all 32warps on fermi issueuing some mul or add, but they have 9tasks on schedule at once, so yeah this totals 1/3 of the flops
<mastermart>
so throughput is not accurate measure of speed, latency is rather that measure, so when the matrix will be made bigger, it would schedule more, hence more throughput be present
<mastermart>
i was also more accurate on the vacant and vfull/stop_fetch variables in queues_not_full assignment in round_robin.v vacant bits go contiguously masking the previous ones, until all are done, then it drives an X if anything can not schedule, which flushes all the queues hence
Rondom has quit [Remote host closed the connection]
Rondom has joined #lima
<mastermart>
if the queue fifo was happen to be full as well, it would drive wr pointer to every warp, and if the queue was 6==6 rd and wr regard
<mastermart>
than it will be 1 and 6 after every warp fetches another instruction
<mastermart>
i.e write pointer wrapper around
maccraft123 has joined #lima
<mastermart>
only difference among cuda and glsl code is, that in GLSL i use specific procedure to schedule stuff, not using cuda nor opencl counter. to restart the threads, but pull the chip in to recursion
<mastermart>
i did go to patent attorney if this can be patented since you are dicks, did that long time ago
<mastermart>
but probably i won't go forward with this one, no interest.
<mastermart>
since i have lost every trial in my home land except one, when people started to pressurise authorities to take sanctions against me, i do not want to deal with such idiots anymore
<mastermart>
like judges are in estonia for instance
<mastermart>
it does not have to be as easy that you can naturally type things eyes closed, but since you have not even come close to sane code, and are flooding the entire os and kernel with you entire crap, i have not even managed to start playing chess yet, i first want to fix the code, but you do nonsense
<mastermart>
chess is also very mental and sometimes depressive game, little similarlities between how queues can me mastered in between the two.
maccraft123 has quit [Quit: WeeChat 2.6]
maccraft123 has joined #lima
<mastermart>
in queues is also, nothing does not work until you become good at mastering the code there with proper deps.
<mastermart>
similarly if you are arrogant dick not knowing anything and going to contend magnus carssen in chess, quick sah mat will be present.
<mastermart>
hence to defend myself, last five years i've been dealing with how queues can be authored in different ways pinned for computing onchip
<mastermart>
and that is what sane people do, they train before excessively
<mastermart>
you can not become grandmaster in chess also without it, and if you would do the same, youd commit less shit
<mastermart>
i do not take it as critisism when you try to bother me, that i have been longer "doing" graphics than new developers have been alive, since lckily none of the crap is pushed by me, all shit is pushed by you
maccraft123 has quit [Quit: WeeChat 2.6]
maccraft123 has joined #lima
<mastermart>
it's very mistaken code that is pushed to the repositories to make it clear, you have not taken stuff seriously, and hence the outcome, graphics enthusiasts over the world are anxious and dissapointed, and channels have very few residents
<mastermart>
maybe vim could do magic on editing the code, it can not type million of lines of retarded apis, wtf. you are broken geniuses in that regard, i think you have allready drained your eyes with this shit, it drains my browser too to see that bulky pointless code
<mastermart>
so that does matter, it matter much negatively, but this irony around me, i practice every day as late go, not to be a shithose like you, thank you! and bye
mastermart has quit [Remote host closed the connection]